What the page is actually saying.

Three widths can appear and they are not interchangeable.

• NHC cone — 2026 operational 2/3-probability circles from 2021–2025 official errors. East Pacific 48 h is 56 n mi; Atlantic 48 h is 62 n mi. Same number every storm in that basin. On by default.
• ECMWF 1σ ellipse — standard deviation of member latitude and longitude around the ensemble mean, drawn as an ellipse (not a circle) on the Helix point at the selected lead. About 68% of the 50 perturbed members plus control sit inside that ellipse. Not a historical error circle, and not the same probability as the official cone.
• Helix cone (off by default) — still last year’s typical Helix miss (holdout p67). We have not replaced it with the swarm. The page says so.

• Default layers — official cone on, cyan 1σ ellipse on, guidance cloud on, Helix cone off, named guidance off. Consensus lines (Helix, OFCL, TVCN, HCCA) stay available; they are not the opening.
• Scrubber — 24, 48, 72, 120 h. It restyles the ellipse at that lead, updates the sidebar sentence, and keeps the same cycle. It does not change the official cone radii; those are still last year’s numbers for that basin and lead.
• Ellipse, not a ring — the overlay uses sd_lat_deg and sd_lon_deg from the decoded 50-member swarm. A circle would lie when the swarm is elongated along or across track.
• Member count on the tooltip — “n of 51.” If members have already dropped the storm, the ellipse is thinning. The sidebar warns below 20 members and refuses to treat it as a full-swarm width below 10.

• Air sentence — the first paragraph. At the selected lead: ECMWF 1σ in n mi, the basin’s official cone, tighter/wider and percent of that cone, intensity ± member σ, and control-versus-ensemble along/cross in words. Plus the member-count warning when it applies.
• Swarm vs cone — the big number is this run’s 1σ. Green/red is that radius minus the 2026 operational cone for the basin. 1σ ≈ 68% of members; the official cone is a historical 2/3 official-error circle.
• Along-track / cross-track — ECMWF control versus ensemble mean, decomposed relative to the control heading. Along is faster/slower (timing). Cross is left/right of heading (location). The small plot is that vector, not the 50-member swarm.
• Intensity — Helix vmax at that lead, plus the ECMWF member standard deviation of vmax when stored. A tight track ellipse with a fat intensity σ is still an uncertain forecast.
• Bifurcation note — if control and mean are farther apart than 75% of the swarm width, or the guidance cloud’s pairwise distances look bimodal, the page says a 1σ ellipse can hide two good solutions. Otherwise it says the cloud looks unimodal. That is a diagnostic, not a cluster analysis of the 50 members.
• Guidance clustering — 0–100 from mean pairwise distance of the interpolated aids Helix used at 48 h. Withheld if fewer than four aids. Not the 50-member swarm.

The table under the map subtracts 1σ from the official cone at every stored lead. Green = this cycle’s swarm is tighter than last year’s typical miss. The bar chart is the same comparison drawn.

Helix is a gradient-boosted residual on a multi-model consensus (v3.0.0). Inputs at advisory time only: public NHC a-deck interpolated aids, ECMWF open-data tracks relocated the way NHC builds EMXI, ECMWF control-versus-ensemble geometry, and SHIPS environment. TVCN is NHC’s track consensus. HCCA is the HFIP corrected consensus. AIFS is shown, not scored, not a trained feature.

Every advisory-cycle run is kept under a cycle-keyed JSON and never overwritten. When the operational b-deck has a fix at the verifying time, Helix computes haversine track error (n mi) and absolute intensity error (kt). Homogeneous columns use only cycles where Helix and all three baselines verified, so the four numbers are comparable. “Helix beat NHC” is the share of individual forecasts where Helix’s track error was smaller. In-season b-decks are operational best tracks, not post-season reanalysis.

• Width — prefer ECMWF 1σ when that cycle stored it. Otherwise the 48 h mean pairwise distance of the aids Helix used (the agreement number). Compare that radius to the NHC cone for that basin and lead. A cycle is “tight” if width < cone, “split” if not.
• The line — homogeneous 48 h track MAE for Helix, OFCL, TVCN, HCCA, plus the share of cycles Helix beat OFCL.
• Different bets — Helix-beat-NHC rate on tight cycles versus split cycles. If those rates are the same, calling the width did not buy a line win. The calibration note under the table is that comparison in one paragraph.

Most 2026 cycles predate ECMWF ingest. Open data has no archive, so those width numbers are guidance spread, not the 50-member swarm. The page says that.

Same verifying protocol as §5, pooled across the season. Homogeneous n is the only fair four-way comparison. The note under the table splits v1 / v2 / v3 counts and ECMWF coverage.

Homogeneous track MAE at each lead — same cycles for Helix, OFCL, TVCN, HCCA. Storm-block bootstrap, 2000 resamples, 95% percentile CI of the mean paired difference (Helix minus baseline). Negative favors Helix. A difference whose interval includes 0 is a tie. Intensity is scored the same way; TVCN has no intensity.